mkn.gpu CUDA/HIP C++20 convenience wrappers ====== Whether you are using CUDA or ROCM, we attempt to deduce from available headers. If automatic detection fails, specify appropriate define like `-DMKN_GPU_CUDA=1` Without CUDA or ROCM headers the CPU backend is used. See: inc/mkn/gpu/def.hpp Compile argument switches Key MKN_GPU_CUDA Type bool Default 1 if cuda_runtime.h is found, else 0 Description activate CUDA as impl of mkn::gpu::* Key MKN_GPU_ROCM Type bool Default 1 if hip/hip_runtime.h is found, else 0 Description activate ROCM as impl of mkn::gpu::* Key MKN_GPU_CPU Type bool Default 1 if neither CUDA nor ROCM is active, else 0 Description activate the CPU backend as impl of mkn::gpu::*, even when GPU headers are available Key MKN_GPU_FN_PER_NS Type bool Default 0 Description expose functions explicitly via mkn::gpu::hip::* mkn::gpu::cuda::* Key MKN_GPU_WARP_SIZE Type uint Default use manufacturer provided (eg warpSize), usually 32 Description override use if defined Key MKN_GPU_TRACE Type bool Default 1 for CUDA when NVTX3 headers are available, else 0 Description profiler annotations (NVTX / roctx), see inc/mkn/gpu/trace.hpp ROCM requires linking rocprofiler-sdk-roctx Key MKN_CPU_DO_NOT_DEFINE_DIM3 Type bool Default false Description if true, skips defining dim3 struct which is usually provided by gpu headers